Hi Tejun,

On 01/03, Tejun Heo wrote:
>
> On Tue, Jan 03, 2012 at 04:44:04PM +0100, Oleg Nesterov wrote:
> > It fails because ->real_parent sees its child in EXIT_DEAD state
> > while the tracer is going to change the state back to EXIT_ZOMBIE
> > in wait_task_zombie().
>
> Argh.... EXIT_ZOMBIE -> DEAD -> ZOMBIE dancing in wait_task_zombie()
> is just nasty.  Didn't realize it was doing that.  :(

We both missed this ;)

> > The offending commit is 823b018e which moved the EXIT_DEAD check,
> > but in fact we should not blame it. The original code was not
> > correct as well because it didn't take ptrace_reparented() into
> > account and because we can't really trust ->ptrace.
> >
> > This patch adds the additional check to close this particular
> > race but it doesn't solve the whole problem. We simply can't
> > rely on ->ptrace in this case, it can be cleared if the tracer
> > is multithreaded by the exiting ->parent.
>
> I'm not following this part.  Can you please explain it in a bit more
> detail?

Before 823b018e the code was:

	if (!ptrace && p->ptrace) {
		wo->notask_error = 0;
		return 0;
	}

	if (p->exit_state == EXIT_DEAD)
		return 0;

There are 2 problems:

	1. it is not correct to clear ->notask_error unless
	   this child is ptrace_reparented(). Nobody will
	   wakeup us if EXIT_DEAD was set by our sub-thread.

	2. We can not rely on ->ptrace to detect this case.

	   Suppose that the tracer is multithreaded, it has
	   two threads T1 and T2, T1 traces our child.

	   - T2 does do_wait(WEXITED), sets EXIT_DEAD, drops
	     tasklist_lock.

	   - T1 exits and does __ptrace_detach(), this means
	     __ptrace_unlink() and nothing more.

	   - Now, real_parent does do_wait() and sees the
	     EXIT_DEAD child but ->ptrace = 0.

	   - finally T2 sets EXIT_DEAD but it is too late,

The patch doesn't solve the 2nd (btw very old) problem. Fortunately
this race is very unlikely.

> > I think we should kill EXIT_DEAD altogether, we should always
> > remove the soon-to-be-reaped child from ->children or at least
> > we should never do the DEAD->ZOMBIE transition. But this is too
> > complex for 3.2.
>
> Agreed.  Removing the reverse transition shouldn't be too difficult
> and can be done without affecting fast non-ptrace path.  ie. if the
> child is ptraced, drop readlock, grab writelock, recheck, buffer
> states to copy out to userland, detach and transit to DEAD if
> necessary.

Yes.

> > Also, I think wait_consider_task() needs more fixes. I do not
> > think we should clear ->notask_error without WEXITED in this
> > case, but this is what we do in the EXIT_ZOMBIE case.
>
> Hmmm... I'm not sure about that.  Why do you think so?

I am not sure too. But why do_wait() should sleep if it is called
without WEXITED (lets ignore WCONTINUED) and the child is ZOMBIE?
I think it should return -ECHILD, like it does if the child is not
traced.

IOW. Suppose we have a single EXIT_ZOMBIE child. If it is not traced,
do_wait(WSTOPPED) returns -ECHILD. If the child is traced (by another
process) do_wait() sleeps until detach just to return the same error.
This looks a bit strange.

>  Acked-by: Tejun Heo <tj@kernel.org>

Great, thanks.

Oleg.